SC-Net: A new strip convolutional network model for rice seedling and weed segmentation in paddy field

Weeds are among the major factors that could harm the yield and quality of rice. Accurately recognizing and localizing crops and weeds are essential for realizing automated weed management in precision agriculture. Semantic segmentation techniques based on deep learning have the capability to automatically discern various types of objects. However, effectively extracting image features to distinguish between rice seedlings and weeds, which often exhibit similar texture characteristics and size disparities, remains a challenging issue in the field. In view of this, a new strip convolutional network model named SC-Net is proposed in this paper, where UNet is used as the backbone network. Based on the idea of multi-scale feature fusion, the parallel multilevel convolution block (PMCB) and strip multilevel convolution block (SMCB) are constructed to design the encoder and decoder of the segmentation network, enabling the extraction of the salient features of seedlings and weeds. Specifically, the SMCB is composed of multi-scale strip convolutions, which effectively widens the receptive field of the convolution layer while minimizing computational costs, and incorporates a long and narrow shape enhancement network to identify the characteristics of slender rice seedling leaves. To adaptively fuse different level features, the attention feature fusion module (AFF) is designed to establish a long skip connection between the encoder and decoder of the network. This module aggregates global and local context information from low-level and high-level features through global spatial pooling and dot product convolution. Moreover, the strip pooling attention module (SPAM) is introduced between the encoder and decoder stages to enhance the network's perception of the precise positional information of the target area, thus further optimizing the segmentation results. The experimental results show that SC-Net achieved MIOU scores of 87.48 % and 89.00 % on the self-built rice seedling and public agricultural datasets. Compared with several state-of-the-art models, the proposed model achieves better segmentation performance, thus contributing to providing a promising support for the development of intelligent weeding in the field.